Abnaolizer: An AI Agent for Converting Antibodies to Nanobodies

ICLR 2026 Conference Submission132 Authors

Keywords: nanobody design, antibody engineering, weakly supervised learning, cross-modal retrieval, Pareto optimization, contrastive learning, protein representation learning, drug discovery

Abstract: Nanobodies, the naturally occurring single-chain antibodies derived from camelids, have emerged as highly promising therapeutic molecules due to their high stability, small size, and ease of engineering. However, generating nanobody candidate sequences from conventional antibodies—one of the primary routes for nanobody development—remains challenging, as rational design is limited by the scarcity of paired data and the complexity of molecular recognition mechanisms. To address this, we propose \textbf{AbNanolizer}, a physics-guided, weakly supervised AI framework for converting conventional antibodies into nanobody candidates. We formalize the task as antigen-conditioned cross-modal retrieval and multi-objective ranking, and design a noise-robust learning scheme to handle weakly paired and mismatched training signals. The framework employs an antigen-conditioned dual-encoder to align sequence representations of conventional antibodies and nanobodies, and jointly optimizes a noise-robust contrastive objective with differentiable Pareto ranking. Optional structural and energetic proxy signals, together with developability predictions, are integrated into a unified optimization. To support reliable decision-making, we perform coverage-guaranteed confidence calibration on retrieval scores. We further construct a rigorous public benchmark and evaluation protocol to enable comparison against strong baselines. Across multiple metrics, AbNanolizer demonstrates consistent improvements and showcases end-to-end applications on three approved drug targets amenable to nanobodies.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 132